ROSE
A framework for lifelong knowledge editing in multimodal large models. It employs orthogonal subspace constraints and risk masking to reduce interference between edits and catastrophic forgetting.
Timeline (last 90 days)
ROSE is a risk-aware orthogonal subspace navigation framework for lifelong knowledge editing in multimodal large models.
ROSE restricts parameter updates to the orthogonal complement of previously edited subspaces, geometrically reducing mutual interference.
The ROSE framework contains two key designs: orthogonal subspace navigation and a risk-guided mechanism, the latter dynamically computing risk masks to protect sensitive parameter subspaces.
ROSE adopts a multi-granularity knowledge integration strategy during inference, including prototype-anchored semantic gating and perturbation-driven subspace fusion.
The researchers conducted comprehensive experiments on ROSE across two mainstream multimodal benchmarks.
ROSE is competitive with current state-of-the-art methods on five key metrics.
ROSE maintains excellent stability even after more than 1,000 sequential updates.
ROSE's code has been open-sourced at https://github.com/sunlitsong/ROSE.
All Facts (8)
ROSE maintains excellent stability even after more than 1,000 sequential updates.
50%UnverifiedROSE is a risk-aware orthogonal subspace navigation framework for lifelong knowledge editing in multimodal large models.
50%UnverifiedThe ROSE framework contains two key designs: orthogonal subspace navigation and a risk-guided mechanism, the latter dynamically computing risk masks to protect sensitive parameter subspaces.
50%UnverifiedROSE adopts a multi-granularity knowledge integration strategy during inference, including prototype-anchored semantic gating and perturbation-driven subspace fusion.
50%UnverifiedROSE restricts parameter updates to the orthogonal complement of previously edited subspaces, geometrically reducing mutual interference.
50%UnverifiedROSE is competitive with current state-of-the-art methods on five key metrics.
50%UnverifiedROSE's code has been open-sourced at https://github.com/sunlitsong/ROSE.
50%UnverifiedThe researchers conducted comprehensive experiments on ROSE across two mainstream multimodal benchmarks.
50%